Start with friction, not with an AI tool
A lot of AI projects begin backwards. Someone sees a new tool, the team schedules a demo, and only then tries to invent a business problem for it. We prefer to start with repeated friction.
For one week, write down the work that happens over and over: copying information from a form into a CRM, rewriting the same type of follow-up, tagging leads, summarizing calls, building the same report, looking for missing fields, routing requests, or turning one approved message into five channel variations. Those are better candidates for automation because the value is measurable: fewer manual steps, faster response times, fewer errors, or more consistent follow-up.
Use AI where judgment is structured
AI is strongest when the task has enough context to make a useful recommendation but still has a clear boundary. For example: summarize an inquiry, classify its intent, propose the next best action, draft a response using approved brand guidance, and then send it to a human for approval.
That is very different from “let AI run marketing.” The first system has a defined input, an expected output, guardrails, and a person who owns the decision. The second is a vague promise that usually creates more review work than it removes.
Separate deterministic automation from AI
Not every step needs intelligence. If a form says the customer is in Ontario, a normal rule can assign the Ontario sales owner. If a customer has not opened a product in 60 days, a normal workflow can trigger a re-engagement path.
Use AI only where the work is genuinely fuzzy: interpreting free text, summarizing, generating variants, extracting patterns, or recommending an action. This keeps the system cheaper, easier to debug, and more predictable.
Build a human checkpoint around expensive mistakes
A useful rule is to ask what happens if the model is wrong. If the result is a draft headline, the risk is low. If the result changes pricing, sends a regulated communication, deletes CRM data, or tells a customer something legally sensitive, the approval layer should be stronger.
The goal is not to remove people from the loop. It is to remove people from low-value repetition so they can spend more time on judgment, creative direction, customer nuance, and decisions that actually deserve attention.
Measure saved work, not AI activity
“We generated 400 AI drafts” is not a business outcome. Better measures are hours saved, time-to-first-response, completion rate, lead-routing accuracy, campaign production time, conversion rate, or the percentage of routine requests handled without manual intervention.
If an AI workflow produces more output but creates extra review, more corrections, or brand inconsistency, it is not automation — it is moving work around.
Our point of view
AI should make the business feel calmer. Fewer tabs. Fewer copy-pastes. Fewer “who owns this?” messages. Faster answers. Better context when a human does need to step in.
We usually recommend automating one contained workflow first, measuring it for a few weeks, and only then expanding. The companies that benefit most from AI are not necessarily the ones using the most tools. They are the ones with the clearest processes underneath them.
Automate the repeatable parts first. Use AI for interpretation and variation, keep humans around consequential decisions, and measure the work the system actually removes.
